Background and Aims

This study investigated whether including endometrial thickness (EMT) could enhance both the predictive accuracy and clinical interpretability of an artificial intelligence (AI) model that uses embryo morphology to forecast in vitro fertilization (IVF) pregnancy outcomes.

Methods

We analyzed 2,016 day-5 fresh embryo transfer cycles from four clinics in South Korea and three in the United States. The dataset included embryo images, maternal age, endometrial thickness (EMT), and pregnancy outcomes based on fetal heart tone (FHT). Four deep learning models were developed using different input combinations, including image only, image with age, image with EMT, and image with both age and EMT and evaluated by macro F1-score and accuracy. Logistic regression and subgroup analyses (EMT ≤ 8 mm) assessed the effects of age and EMT.

Results

Both younger maternal age and thicker EMT were significantly correlated with higher pregnancy success rates (p < 0.05). Compared to the image-only model (F1: 0.55; Accuracy: 61%), incorporating age (F1: 0.59; Accuracy: 68%) or EMT (F1: 0.56; Accuracy: 62%) improved predictive performance. The combined image + age + EMT model achieved similar results to the image + age model (F1: 0.58; Accuracy: 64%). Among high AI-score cases, failed pregnancies exhibited notably thinner EMT than successful ones (p = 0.013). In patients with EMT ≤ 8 mm, adding EMT increased the F1-score from 56.8% to 59.5% and accuracy from 57.9% to 65.8%.

Conclusions

Including EMT in embryo image-based AI models enhances prediction accuracy for patients with thin endometrium and clarifies the cause of false-positive results. Accounting for uterine conditions also improves model transparency and support clinical decision-making in cases of recurrent implantation failure or unexplained IVF outcomes.